Your AI knows the past. Let it see the future.
Credence is a probability API that lets AI agents reason responsibly about the future.
A real decision
Four days before the 2026 Eurovision final, Claude said I should book a venue for my Denmark watch party. Claude + Credence saw that comparable historical forecasts carried thin evidence strength and held back. Denmark finished 7th.
See what happenedYour AI agent has an incomplete view of the future.
Ask your AI to reason about the future and at best it will anchor its views with a prediction-market price — a traded number shaped by fees, liquidity, and trader mix — with no account of how comparable forecasts held up historically.
Credence serves calibrated probabilities and group-conditional evidence strength for “Will X happen?” — so AIs can make better forward-looking decisions for their humans.
One Probability Object. Decision-grade.
Three values, with their scope made explicit.
AIs ask “How likely is event X?” Credence returns our calibrated probability and group-conditional evidence strength, along with the raw market price.
- Calibrated probability — our outcome-calibrated point forecast.
- Evidence strength — how held-out cohorts of comparable forecasts behaved.
- Market price — the captured prediction-market price.
The honest reading
Markets like this one.
Calibration and evidence strength are validated across groups of held-out forecasts. Neither is a promise that one terminal outcome identifies this market’s latent precision.
Three pillars.
Calibration
Refined probabilities of future events, not just market prices.
Evidence Strength
A group-conditional evidence-strength value on every probability, describing the historical cohort assigned to the forecast.
Lineage
Audit-grade source provenance on every response; each number traces to the snapshot, model bundle, and inputs that produced it.
Group-conditional by design
Evidence strength describes the cohort.
It does not claim to recover an unobservable precision parameter for a single market.
Markets contain opaque predictive information.
We serve it cleanly.
Markets are outstanding aggregators of predictive information, but AIs lose a lot of information looking only at prices.
We calibrate prices against outcomes across held-out groups and report evidence strength from comparable historical cohorts.
Built for AI agents first.
Machine-readable by default, with a clean human view on top — for AI agents and anyone who reasons about what happens next.
Available through machine-readable REST and MCP interfaces for agent workflows.
Methodology you can audit.
Every response carries the source timestamp and version identifiers needed to trace how it was produced.
Training is evaluated using walk-forward, out-of-fold backtesting, so each test fold is scored on forecasts not fit to that fold.
Models promoted through formal gates with signed, versioned model bundles, reproducible by digest: every probability object traces back to the model and inputs that produced it.
Built by Jeffrey A. Ryan — Ph.D. in probability (NYU Courant), former US SEC machine-learning lead, former high-frequency trader.